diff --git a/apps/desktop/src/ai/hooks/useLLMConnection.ts b/apps/desktop/src/ai/hooks/useLLMConnection.ts
index 9c2914932e..f4b3507301 100644
--- a/apps/desktop/src/ai/hooks/useLLMConnection.ts
+++ b/apps/desktop/src/ai/hooks/useLLMConnection.ts
@@ -13,6 +13,7 @@ import type { AIProviderStorage } from "@anlg/store";
 
 import { createAppleFoundationModel } from "../apple-foundation-model";
 import { createAuthFetch } from "../auth-fetch";
+import { streamOnlyGenerationMiddleware } from "../stream-only-generation";
 import { createTracedFetch, tracedFetch } from "../traced-fetch";
 
 import { useAuth } from "~/auth";
@@ -291,7 +292,15 @@ const createLanguageModel = (
         baseURL: oauth ? CHATGPT_API_BASE_URL : conn.baseUrl,
         apiKey: oauth ? "oauth" : conn.apiKey,
       });
-      return wrapWithThinkingMiddleware(provider.responses(conn.modelId));
+      const model = provider.responses(conn.modelId);
+      return wrapWithThinkingMiddleware(
+        oauth
+          ? wrapLanguageModel({
+              model,
+              middleware: streamOnlyGenerationMiddleware,
+            })
+          : model,
+      );
     }
 
     case "grok":
diff --git a/apps/desktop/src/ai/stream-only-generation.test.ts b/apps/desktop/src/ai/stream-only-generation.test.ts
new file mode 100644
index 0000000000..05aaedec63
--- /dev/null
+++ b/apps/desktop/src/ai/stream-only-generation.test.ts
@@ -0,0 +1,118 @@
+import { wrapLanguageModel } from "ai";
+import { describe, expect, test, vi } from "vitest";
+
+import { streamOnlyGenerationMiddleware } from "./stream-only-generation";
+
+type LanguageModel = Parameters<typeof wrapLanguageModel>[0]["model"];
+type StreamPart =
+  Awaited<
+    ReturnType<LanguageModel["doStream"]>
+  >["stream"] extends ReadableStream<infer Part>
+    ? Part
+    : never;
+
+const usage = {
+  inputTokens: {
+    total: 4,
+    noCache: 4,
+    cacheRead: 0,
+    cacheWrite: 0,
+  },
+  outputTokens: { total: 2, text: 2, reasoning: 0 },
+};
+
+describe("streamOnlyGenerationMiddleware", () => {
+  test("uses streaming for generate calls and collects the result", async () => {
+    const doGenerate = vi.fn(async () => {
+      throw new Error("non-streaming request used");
+    });
+    const model = createModel(doGenerate, [
+      { type: "stream-start", warnings: [] },
+      {
+        type: "response-metadata",
+        id: "response-1",
+        modelId: "gpt-test",
+        timestamp: new Date("2026-08-25T00:00:00.000Z"),
+      },
+      { type: "text-start", id: "message-1" },
+      { type: "text-delta", id: "message-1", delta: "Hello" },
+      { type: "text-delta", id: "message-1", delta: " world" },
+      { type: "text-end", id: "message-1" },
+      {
+        type: "tool-call",
+        toolCallId: "tool-1",
+        toolName: "lookup",
+        input: '{"query":"test"}',
+      },
+      {
+        type: "finish",
+        finishReason: { unified: "stop", raw: "completed" },
+        usage,
+      },
+    ]);
+    const wrapped = wrapLanguageModel({
+      model,
+      middleware: streamOnlyGenerationMiddleware,
+    });
+
+    const result = await wrapped.doGenerate({ prompt: [] });
+
+    expect(doGenerate).not.toHaveBeenCalled();
+    expect(result.content).toEqual([
+      { type: "text", text: "Hello world" },
+      {
+        type: "tool-call",
+        toolCallId: "tool-1",
+        toolName: "lookup",
+        input: '{"query":"test"}',
+      },
+    ]);
+    expect(result.finishReason).toEqual({ unified: "stop", raw: "completed" });
+    expect(result.usage).toEqual(usage);
+    expect(result.response).toMatchObject({
+      headers: { "x-request-id": "request-1" },
+      id: "response-1",
+      modelId: "gpt-test",
+    });
+  });
+
+  test("rejects provider errors from the response stream", async () => {
+    const model = createModel(vi.fn(), [
+      { type: "stream-start", warnings: [] },
+      { type: "error", error: new Error("upstream failed") },
+    ]);
+    const wrapped = wrapLanguageModel({
+      model,
+      middleware: streamOnlyGenerationMiddleware,
+    });
+
+    await expect(wrapped.doGenerate({ prompt: [] })).rejects.toThrow(
+      "upstream failed",
+    );
+  });
+});
+
+function createModel(
+  doGenerate: LanguageModel["doGenerate"],
+  parts: StreamPart[],
+): LanguageModel {
+  return {
+    specificationVersion: "v3",
+    provider: "test",
+    modelId: "gpt-test",
+    supportedUrls: {},
+    doGenerate,
+    doStream: async () => ({
+      stream: new ReadableStream<StreamPart>({
+        start(controller) {
+          for (const part of parts) {
+            controller.enqueue(part);
+          }
+          controller.close();
+        },
+      }),
+      request: { body: { stream: true } },
+      response: { headers: { "x-request-id": "request-1" } },
+    }),
+  };
+}
diff --git a/apps/desktop/src/ai/stream-only-generation.ts b/apps/desktop/src/ai/stream-only-generation.ts
new file mode 100644
index 0000000000..0c1fcb998b
--- /dev/null
+++ b/apps/desktop/src/ai/stream-only-generation.ts
@@ -0,0 +1,158 @@
+import type { LanguageModelMiddleware } from "ai";
+
+type WrapGenerate = NonNullable<LanguageModelMiddleware["wrapGenerate"]>;
+type GenerateResult = Awaited<
+  ReturnType<Parameters<WrapGenerate>[0]["doGenerate"]>
+>;
+type StreamResult = Awaited<
+  ReturnType<Parameters<WrapGenerate>[0]["doStream"]>
+>;
+type StreamPart =
+  StreamResult["stream"] extends ReadableStream<infer Part> ? Part : never;
+type Content = GenerateResult["content"][number];
+type TextContent = Extract<Content, { type: "reasoning" | "text" }>;
+
+export const streamOnlyGenerationMiddleware: LanguageModelMiddleware = {
+  specificationVersion: "v3",
+  wrapGenerate: async ({ doStream }) => collectStream(await doStream()),
+};
+
+async function collectStream(result: StreamResult): Promise<GenerateResult> {
+  const content: Content[] = [];
+  const openBlocks = new Map<string, TextContent>();
+  let warnings: GenerateResult["warnings"] = [];
+  let responseMetadata: NonNullable<GenerateResult["response"]> = {};
+  let finish: Extract<StreamPart, { type: "finish" }> | undefined;
+
+  const reader = result.stream.getReader();
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+
+      switch (value.type) {
+        case "text-start":
+        case "reasoning-start": {
+          createBlock(value, content, openBlocks);
+          break;
+        }
+        case "text-delta":
+        case "reasoning-delta": {
+          const block = createBlock(value, content, openBlocks);
+          block.text += value.delta;
+          if (value.providerMetadata) {
+            block.providerMetadata = value.providerMetadata;
+          }
+          break;
+        }
+        case "text-end":
+        case "reasoning-end": {
+          const block = createBlock(value, content, openBlocks);
+          if (value.providerMetadata) {
+            block.providerMetadata = value.providerMetadata;
+          }
+          openBlocks.delete(blockKey(value));
+          break;
+        }
+        case "file":
+        case "source":
+        case "tool-approval-request":
+        case "tool-call":
+        case "tool-result":
+          content.push(value);
+          break;
+        case "stream-start":
+          warnings = value.warnings;
+          break;
+        case "response-metadata":
+          responseMetadata = {
+            ...responseMetadata,
+            id: value.id,
+            timestamp: value.timestamp,
+            modelId: value.modelId,
+          };
+          break;
+        case "finish":
+          finish = value;
+          break;
+        case "error":
+          throw value.error;
+        case "raw":
+        case "tool-input-start":
+        case "tool-input-delta":
+        case "tool-input-end":
+          break;
+      }
+    }
+  } finally {
+    reader.releaseLock();
+  }
+
+  if (!finish) {
+    throw new Error("ChatGPT response stream ended without a finish event");
+  }
+
+  const hasResponseMetadata =
+    result.response !== undefined ||
+    responseMetadata.id !== undefined ||
+    responseMetadata.timestamp !== undefined ||
+    responseMetadata.modelId !== undefined;
+
+  return {
+    content,
+    finishReason: finish.finishReason,
+    usage: finish.usage,
+    providerMetadata: finish.providerMetadata,
+    request: result.request,
+    response: hasResponseMetadata
+      ? { ...responseMetadata, ...result.response }
+      : undefined,
+    warnings,
+  };
+}
+
+function createBlock(
+  part: Extract<
+    StreamPart,
+    {
+      type:
+        | "reasoning-delta"
+        | "reasoning-end"
+        | "reasoning-start"
+        | "text-delta"
+        | "text-end"
+        | "text-start";
+    }
+  >,
+  content: Content[],
+  openBlocks: Map<string, TextContent>,
+): TextContent {
+  const key = blockKey(part);
+  const existing = openBlocks.get(key);
+  if (existing) {
+    return existing;
+  }
+
+  const block: TextContent = {
+    type: part.type.startsWith("reasoning") ? "reasoning" : "text",
+    text: "",
+    providerMetadata: part.providerMetadata,
+  };
+  openBlocks.set(key, block);
+  content.push(block);
+  return block;
+}
+
+function blockKey(part: {
+  id: string;
+  type:
+    | "reasoning-delta"
+    | "reasoning-end"
+    | "reasoning-start"
+    | "text-delta"
+    | "text-end"
+    | "text-start";
+}): string {
+  const type = part.type.startsWith("reasoning") ? "reasoning" : "text";
+  return `${type}:${part.id}`;
+}
diff --git a/apps/desktop/src/settings/ai/llm/subscriptions/models.test.ts b/apps/desktop/src/settings/ai/llm/subscriptions/models.test.ts
new file mode 100644
index 0000000000..ab3fd69c4e
--- /dev/null
+++ b/apps/desktop/src/settings/ai/llm/subscriptions/models.test.ts
@@ -0,0 +1,74 @@
+import { Effect } from "effect";
+import { beforeEach, describe, expect, test, vi } from "vitest";
+
+const mocks = vi.hoisted(() => ({
+  fetchJson: vi.fn(),
+  resolveSubscriptionAccess: vi.fn(),
+}));
+
+vi.mock("./access", () => ({
+  resolveSubscriptionAccess: mocks.resolveSubscriptionAccess,
+}));
+
+vi.mock("~/settings/ai/shared/list-common", async (importOriginal) => ({
+  ...(await importOriginal()),
+  fetchJson: mocks.fetchJson,
+}));
+
+import { listSubscriptionModels } from "./models";
+
+describe("ChatGPT subscription models", () => {
+  beforeEach(() => {
+    mocks.fetchJson.mockReset();
+    mocks.resolveSubscriptionAccess.mockReset();
+    mocks.resolveSubscriptionAccess.mockResolvedValue({
+      token: "access-token",
+      credential: { accountId: "account-1" },
+    });
+  });
+
+  test("parses the Codex catalog and omits hidden models", async () => {
+    mocks.fetchJson.mockReturnValue(
+      Effect.succeed({
+        models: [
+          { slug: "gpt-5.6-sol", visibility: "list" },
+          { slug: "codex-auto-review", visibility: "hide" },
+          {
+            slug: "gpt-5.3-codex-spark",
+            visibility: "list",
+            supported_in_api: false,
+          },
+        ],
+      }),
+    );
+
+    await expect(
+      listSubscriptionModels(
+        "chatgpt",
+        "https://api.openai.com/v1",
+        "stored-credential",
+      ),
+    ).resolves.toMatchObject({
+      models: ["gpt-5.6-sol", "gpt-5.3-codex-spark"],
+    });
+    expect(mocks.fetchJson).toHaveBeenCalledWith(
+      "https://chatgpt.com/backend-api/codex/models?client_version=0.145.0",
+      expect.objectContaining({
+        Authorization: "Bearer access-token",
+        "ChatGPT-Account-ID": "account-1",
+      }),
+    );
+  });
+
+  test("does not offer stale fallback models when discovery fails", async () => {
+    mocks.fetchJson.mockReturnValue(Effect.fail(new Error("unavailable")));
+
+    await expect(
+      listSubscriptionModels(
+        "chatgpt",
+        "https://api.openai.com/v1",
+        "stored-credential",
+      ),
+    ).resolves.toEqual({ models: [], ignored: [], metadata: {} });
+  });
+});
diff --git a/apps/desktop/src/settings/ai/llm/subscriptions/models.ts b/apps/desktop/src/settings/ai/llm/subscriptions/models.ts
index 31a025935f..fad0c61f7f 100644
--- a/apps/desktop/src/settings/ai/llm/subscriptions/models.ts
+++ b/apps/desktop/src/settings/ai/llm/subscriptions/models.ts
@@ -24,20 +24,23 @@ import {
 
 const FALLBACK_MODELS: Record<SubscriptionProviderId, string[]> = {
   claude: ["claude-sonnet-4-6", "claude-opus-4-6", "claude-haiku-4-5"],
-  chatgpt: ["gpt-5.3-codex", "gpt-5.4", "gpt-5.2-codex"],
+  chatgpt: [],
   grok: ["grok-4", "grok-4-fast", "grok-3"],
   github_copilot: ["gpt-4.1", "claude-sonnet-4", "gemini-2.5-pro"],
   kimi_code: ["kimi-for-coding"],
 };
 
 const ChatgptModelSchema = Schema.Struct({
-  data: Schema.Array(
+  models: Schema.Array(
     Schema.Struct({
-      id: Schema.String,
+      slug: Schema.String,
+      visibility: Schema.optional(Schema.String),
     }),
   ),
 });
 
+const CHATGPT_CODEX_CLIENT_VERSION = "0.145.0";
+
 const CopilotModelSchema = Schema.Struct({
   data: Schema.Array(
     Schema.Struct({
@@ -129,16 +132,21 @@ async function listChatgptModels(
   }
 
   return pipe(
-    fetchJson(`${endpoint}/models`, headers),
+    fetchJson(
+      `${endpoint}/models?client_version=${CHATGPT_CODEX_CLIENT_VERSION}`,
+      headers,
+    ),
     Effect.andThen((json) => Schema.decodeUnknown(ChatgptModelSchema)(json)),
-    Effect.map(({ data }) => {
-      const models = data.map((model) => model.id);
+    Effect.map(({ models: catalog }) => {
+      const models = catalog
+        .filter((model) => model.visibility !== "hide")
+        .map((model) => model.slug);
       return {
         models,
         ignored: [],
         metadata: extractMetadataMap(
-          data,
-          (model) => model.id,
+          catalog,
+          (model) => model.slug,
           () => ({ input_modalities: ["text", "image"] as const }),
         ),
       };